Hardware Pseudorandom Number Generator for Neural Network Overfitting

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Neural networks face challenges in generating randomness efficiently, particularly in large datasets, where existing methods for implementing randomizing techniques using CPU or GPU are non-trivial and may not fully realize their performance benefits, and overfitting issues arise due to the complexity of networks with millions of parameters.

Innovation Solution

An execution unit for a processor is developed to efficiently generate random values from a predetermined, almost Gaussian distribution using a hardware pseudorandom number generator and normalizing circuitry, which produces bit strings that can be summed to achieve a desired probability distribution, optimizing for hardware efficiency and performance in neural networks.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If CPU or GPU are used to implement randomizing techniques in neural networks, then randomness can be generated, but the performance benefits are not fully realized and the implementation is non-trivial

Engineering Contradiction:
ImproveperformanceVSAvoidimplementation complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent replaces software-based random number generation on CPU/GPU with dedicated hardware circuitry that generates random values directly through electrical signals and logic operations, eliminating the need for complex software implementations and achieving full hardware acceleration

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent introduces a dedicated random value generation unit as an intermediary component between the neural network processor and memory systems, providing random values through a specialized interface that simplifies integration while enabling efficient hardware-based randomness generation

Inventive Principle:
Principle #24Intermediary (Mediator)

2Adaptability or versatility

If large networks with millions of parameters are used, then neural network capability increases, but overfitting occurs where the network remembers training samples rather than extracting relevant features

Engineering Contradiction:
Improvenetwork capabilityVSAvoidgeneralization ability
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent changes the statistical parameters of the random values generated, specifically using non-uniform probability distributions (such as Gaussian or Laplacian distributions) instead of uniform distributions, to provide appropriate regularization effects that prevent overfitting while maintaining network capability

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent implements mechanisms where random values are dynamically introduced during training based on feedback from the network's performance, adjusting the amount and distribution of randomness to optimize both learning capability and generalization

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP3474134B1Generating randomness in neural networks
Publication Date: 2023.09.13 GRAPHCORE LTD
  • EP3474134B1 patent drawingFigure 1
  • EP3474134B1 patent drawingFigure 1A
  • EP3474134B1 patent drawingFigure 2

AI summary

The present relates to invention deals with an execution unit configured to execute a computer program instruction to generate random numbers based on a predetermined probability distribution. The execution unit comprises a hardware pseudorandom number generator configured to generate at least randomised bit string on execution of the instruction and adding circuitry which is configured to receive a number of bit sequences of a predetermined bit length selected from the randomised bit string and to sum them to produce a result.